About ULTRA · Find my patterns
How the compass decides
Two rule tables. Every recommendation on the result page names the rule that fired. Version 0.2, a draft until KU Leuven Augment signs it off.
Worked example
Take a chat assistant with mixed users, judgement content and personal data, started by the user. Rules R3, R4 and R6 fire, so "The user stays in charge" comes first and "The system adapts to the person" second.
Answer 1: which patterns
| Rule | If | Raises | Cue | Source |
|---|---|---|---|---|
| R1 | Users know little about AI, so set expectations and show limits. | The right amount of trust +2 | O1 | IUI 2026, Table 1: low AI literacy raises trust calibration |
| R2 | A wrong answer costs a lot, so show how it works and support oversight. | The system shows how it works +2, The right amount of trust +1 | O2 | IUI 2026, Table 1: high stakes raise transparency and trust |
| R3 | The content is a matter of judgement, so keep the user in charge. | The user stays in charge +2 | O4 | IUI 2026, Table 1: subjective content raises user control |
| R4 | Personal data is involved, so offer opt-out and override, and ask consent before sending where it is needed. | The user stays in charge +2 | O3 | IUI 2026, Table 1: personal data raises user control |
| R5 | The app starts the interaction, so make its actions visible and reversible. | The user stays in charge +1, The system shows how it works +1 | O2, O4 | ULTRA hypothesis v0.2, from the CNEXT and VAIA cases |
| R6 | Your users are mixed, so let them adjust the level of detail. | The system adapts to the person +1 | O5 | IUI 2026, Table 1: diverse users raise personalisation |
The five obligation cues
The cue column points to the five obligation cues used on every pattern card.
- O1 Say what the AI is doing and where its limits are.
- O2 Make human oversight possible, with the evidence a person needs to overrule the system.
- O3 Support people's rights and meaningful control over how their personal data is processed: ask consent where it is needed, and say what is kept.
- O4 Keep the decision and the authorship with the user, and reduce the pull of defaults and automation.
- O5 Accommodate diverse user needs and support equitable access, for example by letting people adjust the level of detail.
Design considerations drawn from the AI Act and the GDPR, from the IUI 2026 paper. Not legal advice or a conformity assessment.
Answer 2: which metrics
| Rule | If | Measure | Source |
|---|---|---|---|
| M1 | Users with little or mixed AI experience: measure satisfaction and intention to use. | they-want-to-keep-using-it | Figure 12: lay users consider satisfaction (lay users of the domain; reading it as AI experience, and adding intention to use, is our translation) |
| M2 | Users with a lot of AI experience: measure trust. | right-amount-of-trust | Figure 12: domain experts consider trust (experts in the domain; reading it as AI experience is our translation) |
| M3 | The AI proposes and a person decides: measure reliance, efficiency, performance and confidence. | following-the-ai, faster-fewer-errors, sure-of-their-decision | Figure 13: decision support |
| M4 | Generative AI writes content: measure information correctness. | catching-made-up-facts | Figure 13: GenAI considers information correctness |
| M5 | Judgement content or personal data, so users need to adjust, override or ask why: measure controllability. | feeling-in-control | Figure 14: interactive explanations consider controllability; the link to judgement content and personal data is our translation (R3, R4) |
| M6 | You are building a dashboard with AI insights: measure understanding of model behaviour. | getting-how-it-decides | Figure 14: global scope considers model understanding; reading a dashboard as a global explanation is our translation |
| M7 | High stakes and dense information: measure cognitive load. | mental-effort | Section 4.12.5: size and structure of the explanation influence cognitive load; the link to high stakes is our translation |
A pattern's score is the sum of the priorities of the goals it serves; ties go to the stronger evidence state, then to the lower pattern number. Design considerations, not legal advice or a conformity assessment.